Research On Recognition Technology Of Unordered Workpieces Grasping Based On Three-dimensional Vision

Using robots to sort and assemble workpieces has become the norm in automated production.In place of humans,robots are gradually becoming the main performers of production activities.However,as the work environments are increasingly complex,most of the targets grabbed by robots have stacking and obscuring problems.That brings a great challenge to the grasping recognition of robots.In order to achieve accurate recognition of unordered workpieces in stacked and obscured states,this paper proposes a two-stage scheme to estimate the pose of workpieces.It can solve the recognition and positioning problems faced by robots in complex industrial environments,and provide a reference for intelligent transformation and upgrading of industrial production.The main research contents of this paper are as follows:1.Construction of vision detection platform and correction of the data.A platform for image data acquisition was built with a light field camera as the core.This platform used the slide rail,lift table and other components in many parts,which can change the camera object distance.This made it suitable for scenarios where targets cannot be captured due to stacking that causes them to be outside the camera’s depth of field.The two-dimensional images and three-dimensional point clouds of targets can be captured and saved through the platform.For the deviation of the acquired data,a correction scheme was designed with the standard gauge block of rank two as the measurement object.The scheme corrected the data from the X,Y and Z axis directions respectively with a scale factor,which was the ratio of the known offsets to the measured offsets in the direction of a single coordinate axis.The data deviation problem was significantly improved after correction,and its average errors in the X,Y and Z directions were1.0968%,0.0503% and 1.8744%,respectively.2.Segmentation and pre-processing of unordered workpieces.In this paper,an algorithm for point cloud segmentation was proposed,which was based on the mapping relationship between two-dimensional images and three-dimensional point clouds.In the algorithm,threshold segmentation,hole filling,morphological opening,connection domain analysis,and mapping relationships were used to segment the workpiece point cloud.The comparison results with DBSCAN and Euclidean clustering segmentation showed that the time spent by the algorithm was only 10.12% and 12.10% of the two clustering algorithms,and the segmentation results of the three were very similar.At the same time,pre-processing process of segmented point clouds was designed in this study,including voxel downsampling,radius filtering and MLS smoothing.Through this process the point cloud data was repaired and reconstructed,as well as the data accuracy was improved.3.Establishment of workpiece feature description and improvement of feature matching algorithm.In this paper,workpiece point cloud models were created by multiview field acquisition,combined with segmentation algorithm.The models were well represented in detail and easy to create.To select the best algorithm for extracting feature points,this study evaluated the repeatability and real-time performance of key point detection algorithms.After that,the FPFH(Fast Point Feature Histogram)was used to build the feature description of point cloud models.An improved matching algorithm of bipartite graph with weights based on KM(Kuhn-Munkres)algorithm was proposed for the problems of many-to-one and poor robustness in traditional KNN(KNearest Neighbor)feature matching.The algorithm adopted fuzzy matching and provided multiple matching paths to find the best solution with highest overall similarity.As shown by the RP(Recall-Precision)curves of alignment,the improved KM algorithm had higher recall and precision in feature matching,and could get more right matching points.4.Accurate estimation of workpiece pose information.This paper completed the point cloud registration using a combined method of coarse and fine registration.Erroneous correspondences in feature matches were removed by geometric constraint relations,and rough poses of the point cloud were calculated from the remaining matches.Then,the exact pose information of workpiece point clouds could be obtained by optimizing the rough pose with ICP(Iterative Closest Point).And the performance of point cloud registration was evaluated with RMSE(Root Mean Square Error).Finally,in order to verify the accuracy of the poses obtained by this study,the offset error and rotation error of the pose-estimation results were calculated.The results showed that the average offset error and the average rotation error of recognition method were 2.944 mm and 1.987°,respectively.It could meet the requirements of industrial grasping for accuracy of pose error.